Working Paper Series
Department of Economics
University of Verona
Granger non-causality tests between (non)renewable energy
consumption and output in Italy since 1861: the (ir)relevance of
structural breaks
Andrea Vaona
WP Number: 19
ISSN:
December 2010
2036-2919 (paper),
2036-4679 (online)
Granger non-causality tests between (non)renewable energy
consumption and output in Italy since 1861: the (ir)relevance of
structural breaks
Andrea Vaona1
Department of Economics Sciences, University of Verona
Palazzina 32 Scienze Economiche
Ex Caserma Passalacqua
Viale dell’Università 3
37129 Verona
E-mail: [email protected]
Kiel Institute for the World Economy
1
The author would like to thank Natalia Magnani for thoughtful conversations. The usual
disclaimer applies.
1
Granger non-causality between (non)renewable energy
consumption and output in Italy since 1861: the (ir)relevance of
structural breaks
Abstract
The present paper considers an Italian dataset with an annual
frequency from 1861 to 2000. It implements Granger non-causality
tests between energy consumption and output contrasting methods
allowing for structural change with those imposing parameter
stability throughout the sample. Though some econometric details
can differ, results have clear policy implications. Energy
conservation policies are likely to hasten an underlying tendency
of the economy towards a more efficient use of fossil fuels. The
abandonment of traditional energy carriers was a positive change.
Keywords: renewable energy, non-renewable energy, real GDP,
Granger-causality, cointegration, structural change.
JEL Classification: C32, Q43.
1
Introduction
The connection between energy consumption and output has been the topic of an extensive
literature surveyed in Lee (2005, 2006), Yoo (2006), Chontanawat et al. (2006) and Payne (2009,
2010a, 2010b). In particular Payne (2009, 2010a) synthesize the often conflicting results obtained
by the literature into four hypothesis. According to the “growth hypothesis”, energy consumption is
a complement of labour and capital in producing output and, as a consequence, it contributes to
growth. The “conservation hypothesis” implies that real GDP might be boosted by a reduction of
energy consumption possibly due to energy conservation policies, aiming at reducing greenhouse
emissions, improving energy efficiency and curtailing energy consumption and waste. If the
“neutrality” hypothesis holds, energy consumption and real output will not have a significant
connection. Finally, the “feedback” hypothesis suggests that more (less) energy consumption results
in increases (decreases) in real GDP, and vice versa.
We follow a very recent stream of literature by distinguishing between renewable and nonrenewable energy consumption (Sari and Soytas, 2004, Ewing et al., 2007, Sari et al., 2008,
Sadorsky, 2009a, Sadorsky, 2009b, Payne, 2009, Payne, 2010c, Apergis and Payne, 2010a, Apergis
and Payne, 2010b, Apergis and Payne, 2010c and Bowden and Payne, 2010). More specifically, we
deepen the research strategy proposed by Payne (2009), that focused on a single country, the US,
rather than on a panel of countries and implemented Granger non-causality tests after Toda and
Yamamoto (1995) over a data sample running from 1949 to 2006. A similar research strategy was
followed by Tsani (2010) for a Greek sample running from 1960 to 2006, though not concerning
the renewable/non-renewable energy consumption dichotomy. Both the studies warn that their
results might be biased by a small sample problem.
We overcome this limitation by analysing a dataset for Italy from 1861 to 2000. Though, according
to Payne (2010a), eleven studies already used Italian data, none of them could rely on a sample with
2
more than 45 observations and none of them2 distinguished between renewable and non-renewable
energy consumption. Furthermore, Italy is highly dependent from energy imports as many other
European countries. Therefore, it well represents a situation where conservation energy policies are
most needed.
Furthermore, after Zachariadis (2007) – where merits and drawbacks of different econometric
approaches are discussed - we do not stop here. We also resort to integration and cointegration
analyses to shed further light on the energy-growth nexus and to assess the robustness of our
results3.
What is more we provide econometric evidence based on estimators allowing for structural break in
the data, not only in unit root testing, but also while performing cointegration and Granger noncausality tests. We do so by adopting the approach by Lütkepohl et al. (2004), that has found scant
applications in the literature on output and energy consumption so far.
The next section illustrates our data. The following one discusses our econometric methods. The
fourth section is devoted to our results and the last one to our conclusions and to policy
implications.
Data description
Our dataset was compiled by Malanima (2006), which also contains a thoughtful discussion
regarding how the series were built and how the Italian energetic system moved over 140 years
from traditional energy sources to modern fossil carriers.
In the present study we consider four variables: the real GDP measured in 1911 prices, nonrenewable energy consumption and two measures of renewable energy consumption. We define
2
With the exception of Sadorsky (2009b), which, however, makes use of panel integration and
cointegration techniques.
3
One further popular approach in the literature relies on autoregressive distributed lag models after
Pesaran et al. (2001). However, we do not believe this approach suits our setting, given that to test
for long-run causality from energy consumption to output one would have to assume that there was
no long-run feedback from output to energy consumption and vice versa (see Pesaran et al., 2001, p.
293). We deem such a-priori assumptions as untenable and, in fact, their assessment should be the
final goal of any research on the energy-output nexus more than its starting point.
3
non-renewable energy consumption (NRE) as the consumption of fossil fuels. On the other hand,
our first measure of renewable energy consumption (RE1) is the one related to hydroelectric,
geothermal, solar and wind power. Note that 96% of RE1 was on average composed by
hydroelectric power. In the second measure of renewable energy consumption (RE2) we also
include traditional energy sources (namely water, wind, animals and firewood). NRE, RE1 and RE2
are all measured in tons of oil equivalent (hereafter toe) and they are set out in Figure 1, while
Figure 2 shows the real GDP. Note that before 1887 RE1 was negligible. Figure 1 clearly
documents the transition from traditional energy carriers to fossil fuels.
To describe our data into some more detail and to allow comparability with other studies, we follow
Tsani (2010) by considering various sub-periods. We take as reference dates the first and second
world wars and the oil price shock of the mid-1970s (Table 1). To start with, it is worth noting that
even in 1861 the consumption of non-renewable energy was rather relevant being above 800,000
toe, especially when compared to hydroelectric power, which was equal to 23 toe in 1887.
Traditional carriers were the main sources of energy, accounting for nearly 7,900,000 toe.
In the period from 1861 to 1918 there was a clear surge of hydroelectric power, as an average real
economic growth of about 2% was accompanied by a growth rate in non-renewable energy
consumption of the order of 3%, by a steady consumption of energy from traditional sources and by
a 32% average rise in consumption of renewable energy. At the same time, manufacturing activities
were taking off, though they were not the first economic sector of the country by number of
employees yet.
Average growth rates from 1919 to 1946 were clearly affected by the 1929 crisis and the second
world war, given that real GDP, NRE and RE2 all decreased. However, hydroelectric power
continued its rise growing on average by approximately 4% a year.
The following years - especially the 1950s - are renown as those of the “Italian miracle”, thanks to
which Italy managed to catch up with the most developed countries. During this period Italy
completed its industrialization and tertiarization processes as employees in manufacturing and
4
services eventually outnumbered those in agricultural activities and emigration progressively shrank
(Zamagni, 1990, p. 50). The real GDP grew on average by more than 6% a year. This leap was
achieved relying more on non-renewable energy consumption - which increased by nearly 13% a
year – than on RE1 or RE2 – whose growth rates were much more muted.
Finally after the oil price shock of the mid-seventies, the Italian economy was unable to grow as
fast as before, notwithstanding its eventual membership to the European Monetary Union, and
energy consumption indicators mirrored this slower trend, which was also accompanied by a
ballooning public debt and by increasing difficulties to have a positive trade balance. The fact that
energy consumption was growing at a slower pace, however, was accompanied by the increasing
weight of energy imports: energy dependence, measured by net imports divided by the sum of gross
inland energy consumption plus bunkers, passed from 0.4% in 1972 to 78.6% in 2000 and to 85%
in 2004 – similar figures to those showed by Tsani (2010) for Belgium, Greece, Ireland,
Luxembourg, Portugal and Spain. In 2004, fossil fuels were the sources of 87% of energy
consumption in Italy, also as a result of a referendum against nuclear energy production in 1987.
The country under analysis is, therefore, not only far from the condition of the UK, which is a net
energy exporter, but also from those of Germany and France, that can rely on nuclear power and the
former also on carbon (Bastianelli, 2006).
These facts - together with the obligation of reducing CO2 emissions by 6.5% between 2008 and
2012 following commitments to the Kyoto protocol (IEA, 2009), with shrinking oil reserves and
increasing world population - give energy conservation a high priority in the Italian political
agenda. As a consequence, better understanding the connection between energy consumption and
economic growth has an ever growing importance.
5
Methodology
Under a methodological point of view, we contrast the results of econometric tests and estimators
allowing for structural breaks in the data with those imposing parameter stability across the whole
sample.
Methods allowing for structural breaks
In the first case, we test for the presence of a unit root in the series under analysis following Perron
(1989) and Zivot and Andrews (1992). More specifically we adopt Perron's model C
( ) + αy
*
yt = µ + θDU t + β t + γDTt + dD Tτ
t
k
t − 1 + Â ci ∆yt − i + ε t
i =1
where 1 < Tk < T assigns the break point, D(Tk) = 1 if t = Tk + 1 and 0 otherwise, DUt = 1 for t > Tk
and 0 otherwise, DT*t = t & Tk for t > Tk and 0 otherwise. yt is the t-th observation of the variable
under analysis, µ, θ, β, γ, α, d, ci for i=1,…, k are parameters to be estimated, ∆ is the first
difference operator and ε is a stochastic error. Following Zivot and Andrews (1992) we choose the
date of the structural break as the point in time for which the null hypothesis of a random walk with
drift is most likely not to be accepted. The test statistic is the Student t ratio
[ ]
tα λinf = inf tα (λ )
λ ∈Λ
where Λ is [0.15, 0.85] and λ= (Tk/T). As it is possible to see this test can capture a change both in
the mean and in the trend of a given series. Note that k was selected by means of a Schwarz
criterion starting from a maximum lag number of 8.
We run this test for variables both in levels and in first differences. Our final target is to run
Granger non-causality tests on bivariate VARs or VECMs of real output and one energy
consumption measure, so we consider real GDP and one energy consumption measure at a time.
If we find that both these variables are I(1), we will move to a cointegration test. If not, we will
check whether first differenced variables are all stationary. If it is so and if we do not find evidence
6
of cointegration, we will test for short run causality by adopting the Box and Jenkins (1970)
approach. In other words, we will differentiate variables to estimate a stationary VAR and use
customary Granger causality tests, without fear of incurring in possible omitted variable biases due
to the omission of the error correction part of the model4, given that variables do not appear to be
related in the long-run. On the other hand, if we find evidence of the existence of cointegration
relationships we will estimate a vector error correction model (VECM).
At this stage of our research we will take into consideration the possible impact of structural breaks
on our estimates as well. Once adopting the Box and Jenkins approach, we will test for the presence
of structural breaks by resorting both to Quandt and Andrews tests and to a CUSUM test. Once
estimating a VECM, instead, we will follow the procedure proposed by Lütkepohl et al. (2004) as
implemented by Pfaff (2008).
Lütkepohl et al. (2004) consider a (K×1) vector process {yt} generated by a constant (µ0), a linear
trend (t), and level shift terms
yt = µ0 + µ1t + hdtk + xt
where bold characters denote vectors, µ1 is the vector of coefficients of the time trend, dtk is a
dummy variable with dtk = 0 for t < k and dtk = 1 for t œ k, h is the vector of coefficients of dtk. The
shift point k is assumed to be unknown and it is expressed as a fixed fraction of the sample size,
k = [T ] with 0 <
where
0
and
1
0
ø ø
1
<1
define real numbers and [·] defines the integer part. xt is assumed to be
representable by a VAR of order p and to have components that are at most I(1) and cointegrated
with rank r. Note that µ1 could be 0. After Trenkler (2003), the break point is selected on the basis
of the estimation of a VAR(p) in levels for the variable yt, where it is possible to include or not to
include a time trend or seasonal dummies. At this stage, the data are adjusted according to
xˆ t = yt − ˆ 0 − ˆ 1t − ˆd t
4
On this point see for instance Davidson et al. (1978).
7
A Johansen-type test for determining the cointegration rank can be applied to these adjusted series.
If the existence of cointegration is not rejected, we will conduct Granger non-causality tests on the
VECM of the adjusted series. Note that as a first step we will test for the presence of a time trend in
the VAR in levels for {yt}, to select the most suitable model specification.
Methods imposing parameter stability throughout the sample
Given that most of the literature on energy consumption and growth imposes stability of the
estimated parameters, we are curious to see how the results obtained with the methods above
compare with more traditional estimation techniques as this could lead to guidance for future
research.
Similarly to Payne (2009) and Tsani (2010) we follow the approach proposed by Toda and
Yamamoto (1995) as popularized by Rambaldi and Doran (1996) and Zapata and Rambaldi (1997).
This approach to Granger non-causality is as follows. Consider bivariate VAR models between the
logs of the level of real GDP (Y) and each one of our three measures of energy consumption (E).
k
k + d max
k
k + d max
m =1
j = k +1
m =1
j = k +1
E t = a0 + Â a1m E t − m +
 a2 j Et − j +  γ 1mYt −m +
k
k + d max
m =1
j = k +1
Yt = b0 + Â b1mYt − m +
Âb
Y
2 j t− j
Âγ
k
k + d max
m =1
j = k +1
+ Â δ 1m E t − m +
Âδ
2j
Y
2 j t− j
+ ε 1t
Et − j + ε 2 t
where E equals NRE or RE1 or RE2, a0, a1m, a2j, γ1m, γ2j, b0, b1m, b2j, δ1m, δ2j are parameters to be
estimated and ε1t and εtj are disturbances. We choose k by resorting to the Schwarz information
criterion and we set dmax equal to the maximum suspected order of integration of our data series.
Similarly to Tsani (2010), we try to detect it by running a battery of unit root and stationarity tests,
such as the Augemented Dickey Fuller test (after Dickey and Fuller, 1979), the Phillips and Perron
(1988) test and the Kwiatkowski et al. (1992) test, hereafter labelled KPSS. We run all the tests
both with and without a time trend.
Afterwards, Granger non-causality is tested for by means of Wald tests focusing only on the
coefficients γ1m and δ1m for m=1,…,k. Unidirectional Granger causality from real GDP to the energy
8
consumption measure cannot be rejected if γ1mŒ0 for all m. Conversely, unidirectional Granger
causality from energy consumption to real GDP cannot be rejected if δ1mŒ0 for all m. Bidirectional
Granger causality cannot be rejected if γ1mŒ0 and δ1mŒ0 for all m. Interpreting the result in the light
of the “conservation”, “growth” or “feedback” hypotheses will necessitate to take into account also
the sign of the estimated coefficients. Finally, if we can impose the restriction γ1m=0 and δ1m=0 for
all m, we will interpret it as supporting the “neutrality” hypothesis. Note that this procedure has
only asymptotic validity, therefore it is not properly suitable for testing for structural breaks, as one
might wonder whether the results of such tests are due to finite sample distortions more than to the
presence of real structural changes5.
We complement our analysis above also adopting the cointegration tests after Johansen (1991,
1995) without following the Lütkepohl et al. (2004) procedure. On the basis of unit root and
stationarity tests not allowing for a structural break, we check whether the variables under scrutiny
have the same order of integration. If there is convincing evidence that the variables are I(1), we
will specify a vector error correction model (VECM) and test for cointegration. If we find evidence
of cointegration, we will perform causality tests on the VECM. If we do not find evidence of either
cointegration or of the same order of integration in the variables, but we find that first differenced
variables are all stationary, we will resort to the Box and Jenkins (1970) approach.
Results
Methods allowing for structural breaks
Table 2 shows the results of our unit root tests after Perron (1989) and Zivot and Andrews (1992).
A clear pattern emerges. The logs of the real GDP and of NRE are I(1), while those of RE1 and
RE2 are I(0). First differencing always produces stationary variables.
5
At any rate, for sake of completeness, we report in the Appendix our results obtained a rolling regression technique
within a Toda-Yamamoto approach.
9
As a consequence, we proceed with cointegration testing between the former two variables.
Regarding the other two, instead, we specify a VAR in first differences and test for Granger noncausality within this setup.
Regarding the logs of the real GDP and of NRE, we first specify a VAR in levels to choose by
means of the Schwarz information criterion the most suitable number of lags, which turns out to be
two. Furthermore a linear trend would not appear to have a significant coefficient once inserted in
the model, having t-statistics equal to 1.07 and 1.81 in the equations for the log of the real GDP and
the log of NRE respectively. As a consequence, we specify a VECM with one lag in the first
differences and one lag in levels, without any trend.
In this setting, the Lütkepohl et al. (2004) test for cointegration can reject the null of no
cointegration, returning a statistic equal to 17.98 in face of a 1% critical value of 16.42. It can also
reject the null that there does not exist one cointegration relation as it returns a statistic equal to
5.55, larger than the 5% critical value of 4.12. The break point is found to be in 1947.
As a consequence we can proceed with Granger non-causality tests on the data transformed à la
~
~
Lütkepohl et al. (2004), that we denote Yt and NR Et for the logs of the real GDP and of non-
renewable energy consumption. Equations 1 and 2 show our VECM estimates. T-statistics are
reported in brackets below the relevant coefficient.6
∆Yt = − 0.13 ∆NR Et −1 + 0.30 ∆Yt −1 − 0.22ÃÄ Yt −1 − 0.43 NR Et −1 ÔÕ + u1t
(1)
∆NR Et = − 0.21 ∆NR Et −1 + 0.22 ∆Yt −1 − 0.34 ÃÄ Yt −1 − 0.43 NR Et −1 ÔÕ + u2 t
(2)
~
~
~
[− 1.26 ]
~
~
[− 1.30 ]
~
~
[0.85 ]
~
[− 3.63 ]Å
[1.87 ]
~
[− 3.58 ]Å
Ö
[−7.50 ]
~
[−7.50 ]
Ö
where u1t and u2t are disturbances. Short-run coefficients do not appear to be significantly different
from zero. On the contrary it appears that there exists a negative long-run relationship between
energy consumption and output. Short-run dynamics is dominated by adjustment towards the long
run equilibrium, whereby if, for instance, there is a positive deviation of output from its long-run
6
Note that Quandt and Andrews tests as well as CUSUM test applied to equations (1) and (2) would not find any
evidence of structural breaks.
10
relationship with non-renewable energy consumption the growth rates of both variables will
~
decline, though ∆NR Et at a faster speed. This implies that there exists bidirectional Granger
causality between the two variables under study. A greater non-renewable energy consumption
boosts the growth rate of output, but an increase in output depresses the growth rate of nonrenewable energy consumption. The latter effect is stronger than the former one and it could be due
to the fact that economic growth is accompanied by greater efficiency in energy use (on this point
see for instance Huang et al., 2008).
Let us now move to consider the link between the logs of the real GDP and renewable energy
consumption measures, RE1 and RE2. As mentioned above we estimate bivariate VARs in
differences, whose lag orders are set to two on the basis of the Schwarz criterion.
Regarding the VAR model of real GDP and RE1 in logs, in the equation of RE1 a linear trend is
found to have a coefficient significantly different from zero and, on the basis of Quandt and
Andrews unknown breakpoint tests – after Andrews (1993) and Andrews and Ploberger (1994) and of CUSUM tests, two mean shifts are found in the model, the former in 1956 and the latter
1991. The first two rows of Table 3 show the results of Granger non-causality tests within this
model, once adopting a seemingly unrelated estimator. The null cannot be rejected in either
direction, which would favour the neutrality hypothesis between RE1 and real GDP.
Regarding the VAR model of real GDP and RE2 in logs, instead, a linear trend is not found to have
a coefficient significantly different from zero. Furthermore, Quandt and Andrews unknown
breakpoint tests and CUSUM tests cannot find any evidence of structural breaks. On these grounds
a simple VAR(2) model is estimated. The second two rows of Table 3 show the results of Granger
non-causality tests within this model, once adopting a seemingly unrelated estimator. Unidirectional
causality runs from RE2 to real GDP with a negative sign. The following section illustrates our
results once adopting methods that do not allow for structural breaks.
11
Methods imposing parameter stability throughout the sample
We start with the Toda and Yamamoto (1995) approach. Table 4 shows the results of our unit root
and stationarity tests. As it is possible to see the maximum detected order of integration is one for
real GDP, NRE and RE2, while for RE1 the KPSS test would point to two. On the basis of the
results of the Schwarz information criteria mentioned above, we choose two lags for all the three
VARs considered. So we estimate two bivariate VAR(3) models for the logs of real GDP and NRE
and for those of real GDP and RE2 respectively. For real GDP and RE1, instead, we estimate a
VAR(4).
The results of Granger non-causality tests are set out in Table 57. Granger non-causality is rejected
from real GDP to NRE, from NRE to real GDP and from real GDP to RE2.
k
It is worth noting that for non-renewable energy consumption
Â γˆ 1m < 0 and
m =1
k
 δˆ
1m
> 0 , where
m =1
γ̂ 1m and δ̂ 1m are the estimated counterparts of γ 1m and δ 1m respectively. So, similarly to the case
for the VECM with structural breaks, greater non-renewable energy consumption boosts output, but
an increase in output depresses non-renewable energy consumption.
Once including in our renewable energy measure traditional sources of energy, instead, we find unidirectional Granger causality from RE2 to real GDP. This is hardly surprising given that economic
development in Italy, as elsewhere, was characterized by a transition from traditional energy
sources to fossil fuels. Finally, we cannot reject Granger non-causality in either direction for RE1.
Regarding integration and cointegration analyses, we first note that on the basis of the unit root and
stationarity tests above it is not possible to understand whether RE1 is either I(0) or I(2). In either
case, its integration order would appear to be different than the one of real GDP, for which there is
rather clear evidence to be I(1). To estimate a stationary VAR one should, therefore, include
7
Note that we carried out a Portmanteau autocorrelation test for the residuals of all the estimated
VARs without finding any evidence of serial correlation.
12
variables with inhomogeneous difference orders, for which it is difficult to provide an economic
intuition. We conclude that RE1 and real GDP are not connected without any further testing.
Regarding NRE and RE2 - for which there is stronger evidence to be I(1) - we specify the following
VECM on the basis of the Schwarz criteria mentioned above and similarly to Zachariadis (2007):
∆Et = c01 + c11 ∆Et −1 + c 21 ∆Yt −1 + c31 (Et −1 + d 11Yt −1 + d 01 ) + v1t
∆Yt = c02 + c12 ∆Et −1 + c 22 ∆Yt −1 + c32 (Et −1 + d 11Yt −1 + d 01 ) + v 2 t
where cij for i=0,…,3 and j=1,2 and dlj for l=0,1 and j=1,2 are parameters to be estimated, v1t and v2t
are disturbances and E equals either NRE or RE2. In this framework, we implement the Johansen
(1991, 1995) tests, which, however, do not find any evidence of cointegration as showed in Tables
6 and 7.
As a consequence we abandon the VEC model and we test for short-run Granger causality by
specifying two bivariate VAR models in first differences for Y and NRE and for Y and RE2
respectively. We choose a lag length of 2 resorting to the Schwarz information criterion8.
For non-renewable energy consumption our results very closely resemble those obtained with the
Toda and Yamamoto (1995) approach, as Granger non-causality can be rejected both from NRE to
Y and viceversa. In the first case, the Wald statistic is equal to 7.62 with a p-value of 0.02 and the
sum of the coefficients of the lags of NRE in logs is 0.11. In the second case, the Wald statistic is
equal to 8.01 with a p-value of 0.01 and the sum of the coefficients of the lags of real GDP in logs
is -0.88.
For RE2, instead, the results coincide with those presented in Table 3, given that we did not find
evidence of structural breaks.
8
Also in these models the Portmanteau autocorrelation test would not find any evidence of serial
correlation in the residuals of the estimated VAR models.
13
Conclusions
In this paper we contrasted econometric methods allowing for structural breaks with those imposing
parameter stability regarding the issue of the energy consumption-output nexus, distinguishing
between renewable and non-renewable energy sources and using Italian data since 1861. Table 8
offers a summary of our results.
Under an econometric point of view, using methods allowing for structural change can produce in
some cases different results than using methods imposing parameter stability. For instance, the
Lütkepohl et al. (2004) test finds a cointegrating relationship between NRE and real GDP, that
standard Johansen (1991, 1995) tests cannot find. Other examples concern the maximum detected
integration order of RE1 and RE2 in logs. The Zivot and Andrews (1992) test finds both variables
to be I(0), while adopting tests imposing parameter stability one can find a maximum integration
order of 2 and 1 respectively.
When it comes, however, to Granger non-causality tests, results are stable across different
methodologies. We find evidence pointing to bi-directional causality between non-renewable
energy consumption and output, whereby a greater non-renewable energy consumption boosts
output growth, but an increase in the level of output depresses the growth rate of non-renewable
energy consumption to a larger extent, possibly due to greater efficiency in energy use. Granger
non-causality could not be rejected from renewable energy consumption to output and viceversa.
Once including in our renewable energy measure traditional energy sources, negative causality runs
from energy consumption to output.
These results have clear policy implications. Given the prevailing negative effect of output on nonrenewable energy consumption, one can think that the economy tends to make a more efficient use
of this kind of energy as time passes. Conservation policies favouring energy saving in buildings,
lighting and transportation can be thought to hasten an underlying tendency of the economy and
they should be pursued notwithstanding a possible small negative impact on output. All the more
that such an impact takes place only in the short run according to cointegration analysis. Our second
14
result points to a possible limitation to the research strategy inaugurated by Payne (2009) and that
we adopted, as Granger non-causality tests can only give us a retrospective knowledge on the
dynamics of renewable energy consumption and output. Given that in the past, renewable energy
consumption in Italy was mostly connected to hydroelectric power, we might not be able to detect if
we were on the eve of a transition from fossil fuels to geothermal, aeolian or solar power.
This implies that our results cannot provide policy advice regarding more recent alternative energy
sources. They only point to the fact that hydroelectric power cannot completely substitute for fossil
fuels.
Finally, it appears that policies favouring the abandonment of traditional energy carriers were well
suited as their use is negatively associated with output.
One further limitation of the present work, which is common to a great number of studies in the
field, is that, though we adopted a much larger sample than the bulk of the literature, we could
estimate only bivariate models due to historical data availability.
References
Andrews, Donald W. K. (1993). “Tests for Parameter Instability and Structural Change With
Unknown Change Point,” Econometrica, 61(4), 821–856.
Andrews, Donald W. K. and W. Ploberger (1994). “Optimal Tests When a Nuisance
Parameter is Present Only Under the Alternative,” Econometrica, 62(6), 1383–1414.
Apergis N. and J.E. Payne, Renewable energy consumption and economic growth: evidence
from a panel of OECD countries, Energy Policy 38 (2010a), pp. 656–660.
Apergis Nicholas, James E. Payne, Renewable energy consumption and growth in Eurasia,
Energy Economics, Volume 32, Issue 6, November 2010c, Pages 1392-1397, ISSN 0140-9883,
DOI: 10.1016/j.eneco.2010.06.001.
Apergis, N. and J.E. Payne, “The Renewable Energy Consumption-Growth Nexus in Central
America”, Working Paper (2010b).
15
Bastianelli, F. (2006), La politica energetica dell’Unione Europea e la situazione dell’Italia,
La comunità internazionale, 61, pp. 443-468.
Bowden, N. and J.E. Payne, “Sectoral Analysis of the Causal Relationship between
Renewable and Non-Renewable Energy Consumption and Real Output in the U.S.”, Energy
Sources, Part B:, Economics, Planning, and Policy 5 (4) (2010), pp. 400–408.
Box, G.E.P., and G. M. Jenkins, (1970), Time series analysis: forecasting and control,
Holden Day: San Francisco.
Chontanawat, J., Hunt, L.C., Pierse, R., 2006. Causality between energy consumption and
GDP: evidence from30 OECD and 78 non-OECD countries. Paper No. SEEDS 113. Surrey Energy
Economics Discussion Paper Series. University of Surrey.
Davidson, J.E.H., D. F. Hendry, F. Srba, and S. Yeo, (1978), “Econometric modelling of the
aggregate time series relationships between consumer’s expenditure and income in the United
Kingdom”, Economic Journal 88, 661-92.
Dickey, D.A. and W.A. Fuller, Distribution of the estimators for autoregressive time series
with a unit root. Journal of the American Statistical Society 75 (1971), 427-431.
Ewing B.T., R. Sari and U. Soytas, Disaggregate energy consumption and industrial output
in the United States, Energy Policy 35 (2007), pp. 1274–1281.
Huang, B.N., Hwang, M.J. and Yang, C.W. (2008), “Causal relationship between energy
consumption and GDP growth revisited: a dynamic panel data approach”, Ecological Economics,
Vol. 67, pp. 41-54.
IEA (2009), Energy Policies of Italy -2009 Review, OECD-International Energy Agency,
Paris.
Johansen, Søren (1991). “Estimation and Hypothesis Testing of Cointegration Vectors in
Gaussian Vector Autoregressive Models,” Econometrica, 59, 1551–1580.
Johansen, Søren (1995). Likelihood-based Inference in Cointegrated Vector Autoregressive
Models, Oxford: Oxford University Press.
16
Kwiatkowski, D., P.C.B. Phillips, P. Schmidt and Y. Shin, Testing the null hypothesis of
stationarity against the alternative of a unit root: how sure are we that economic time series have a
unit root?, Journal of Econometrics 54 (1992), pp. 159–178.
Lee, C.-C., 2005. Energy consumption and GDP in developing countries: a cointegrated
panel analysis. Energy Economics 27 (3), 415–427.
Lee, C.-C., 2006. The causality relationship between energy consumption and GDP in G-11
countries revisited. Energy Policy 34, 1086–1093.
Lütkepohl, H., Saikkonen, P. and Trenkler, C. (2004), Testing for the cointegrating rank of a
VAR with level shift at unknown time, Econometrica 72, 647–662.
MacKinnon, James G. (1996). “Numerical Distribution Functions for Unit Root and
Cointegration Tests,” Journal of Applied Econometrics, 11, 601-618.
MacKinnon, James G., Alfred A. Haug, and Leo Michelis (1999), “Numerical Distribution
Functions of Likelihood Ratio Tests for Cointegration,” Journal of Applied Econometrics, 14, 563577.
Malanima, P. (2006), Energy consumption in Italy in the 19th and 20th Century, Consiglio
Nazionale delle ricerche, Istituto di Studi sulle Società del Mediterraneo.
Payne, James E., On the dynamics of energy consumption and output in the US, Applied
Energy, Volume 86, Issue 4, April 2009, Pages 575-577.
Payne, J.E., 2010a. Survey of the international evidence on the causal relationship between
energy consumption and growth. Journal of Economic Studies 37, 53–95.
Payne, J.E., 2010b. A survey of the electricity consumption-growth literature. Applied
Energy 87, 723–731.
Payne, J.E. (2010c), “On Biomass Energy Consumption and Real Output in the U.S.”,
Energy Sources, Part B: Economics, Planning, and Policy, forthcoming.
Perron, P., 1989. The great crash, the oil price shock and the unit root hypothesis.
Econometrica 57 (6), 1361–1401.
17
Pesaran, M.H., Shin, Y., Smith, R.J., 2001. Bounds testing approaches to the analysis of
level relationships. Journal of Applied Econometrics 16, 289–326.
Pfaff, B. (2008), Analysis of Integrated and Cointegrated Time Series with R, Spinger, New
York, USA.
Phillips, P.C.B. and P. Perron, Testing for a unit root in time series regression, Biometrika
75 (1988), pp. 335–346.
Rambaldi, A.N. and H.E. Doran (1996), "Testing for Granger Non-Causality in Cointegrated
Systems Made Easy", Working Papers in Econometrics and Applied Statistics- Department of
Econometrics. The University of New England, No. 88, 22 pages.
Sadorsky, P. Renewable energy consumption and income in emerging economies, Energy
Policy 37 (2009a), pp. 4021–4028.
Sadorsky, P. Renewable energy consumption, CO2 emissions and oil prices in G7 countries,
Energy Economics 31 (2009b), pp. 456–462.
Sari, R. and U. Soytas, Disaggregate energy consumption, employment, and income in
Turkey, Energy Economics 26 (2004), pp. 335–344.
Sari, R., B.T. Ewing and U. Soytas, The relationship between disaggregate energy
consumption and industrial production in the United States: an ARDL approach, Energy Economics
30 (2008), pp. 2302–2313.
Toda, Hiro Y. and Taku Yamamoto, Statistical inference in vector autoregressions with
possibly integrated processes, Journal of Econometrics, Volume 66, Issues 1-2, March-April 1995,
Pages 225-250.
Trenkler, C. [2003], ‘A new set of critical values for systems cointegration tests with a prior
adjustment for deterministic terms’, Economics Bulletin 3(11), 1–9.
Tsani Stela Z., Energy consumption and economic growth: A causality analysis for Greece,
Energy Economics, Volume 32, Issue 3, May 2010, Pages 582-590, ISSN 0140-9883, DOI:
10.1016/j.eneco.2009.09.007.
18
Yoo, S.-H., 2006. The causal relationship between electricity consumption and economic
growth in the ASEAN countries. Energy Policy 34 (18), 3573–3582.
Zachariadis, T. Exploring the relationship between energy use and economic growth with
bivariate models: New evidence from G-7 countries (2007) Energy Economics, 29 (6), pp. 12331253.
Zamagni, V. (1993), Dalla periferia al centro. Il Mulino, Bologna.
Zapata, Hector O & Rambaldi, Alicia N, 1997. "Monte Carlo Evidence on Cointegration and
Causation," Oxford Bulletin of Economics and Statistics, Department of Economics, University of
Oxford, vol. 59(2), pages 285-98, May.
Zivot, E., Andrews, D.W.K., 1992. Further evidence on the Great Crash, the oil price shock
and the unit root hypothesis. Journal of Business and Economic Statistics 10 (3), 251–270.
19
Appendix
The present appendix illustrates our results obtained by a rolling regression technique applied to the
three bivariate VARs between real GDP on one side and NRE, RE1, and RE2 on the other - as
specified at p. 12. We discuss these results in an appendix because the Toda and Yamamoto (1995)
approach has only asymptotic validity and reducing the number of observations might increase
finite sample biases.
For the VARs between real GDP and NRE and between real GDP and RE2 we use a window width
of 100 observations. Instead, for the VAR between real GDP and RE1, having less observations, we
use a window width of 90. These widths were chosen in the attempt to ward off the above
mentioned risk of finite sample biases.
The continuous lines in Figures A1 to A3 represent the sums of the lagged coefficients (similarly to
the fifth column of Table 5), while the dotted lines the p-values of modified Wald statistics (like in
the fourth column of Table 5).
As it is possible to see results regarding non-renewable energy consumption are remarkably stable.
Concerning renewable energy consumption 1, some Granger causality running from the real GDP
to RE1 shows up in earlier samples. However, the magnitude of the sums of the coefficients is so
small to be negligible under an economic point of view. Finally, Figure A3 confirms negative
Granger causality running from RE2 to real GDP. In earlier samples, an increase in real GDP
significantly Granger causes an increase in RE2. However, after the sample running from 1878 to
1977, such effect vanishes, as the transition from traditional energy carriers to fossil fuels began to
take place.
20
Figure 1 – Non-renewable and renewable energy consumption in Italy from 1861 to 2000
180,000,000
14,000,000
160,000,000
NRE (left axis)
140,000,000
RE2 (right axis)
12,000,000
10,000,000
RE1 (right axis)
120,000,000
100,000,000
8,000,000
80,000,000
6,000,000
60,000,000
4,000,000
40,000,000
2,000,000
20,000,000
1996
1991
1986
1981
1976
1971
1966
1961
1956
1951
1946
1941
1936
1931
1926
1921
1916
1911
1906
1901
1896
1891
1886
1881
1876
1871
1866
0
1861
0
Notes:
1. NRE is the consumption of fossil fuels. RE1 is energy consumption related to hydroelectric
power, geothermal, solar and wind. RE2 includes RE1 and traditional energy sources (like
water, wind, animals and firewood). Data for RE1 is from 1887.
2. All data are in tons of oil equivalent.
Figure 2 – Real GDP in Italy from 1861 to 2000 (in 1911 prices)
Real GDP
300,000,000
250,000,000
200,000,000
150,000,000
100,000,000
50,000,000
1996
1991
1986
1981
1976
1971
1966
1961
1956
1951
1946
1941
1936
1931
1926
1921
1916
1911
1906
1901
1896
1891
1886
1881
1876
1871
1866
1861
0
21
Figure A1 – Rolling regression Granger non-causality tests (Toda and Yamamoto approach)
From non-renewable energy consumption to real GDP
0.06
0.05
0.04
sum of lagged coefficients
p-values of the modified Wald statistic
0.03
0.02
0.01
0
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
31
33
35
37
From GDP to non-renewable energy consumption
0.4
0.2
0
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
-0.2
sum of lagged coefficients
p-values of the modified Wald statistic
-0.4
-0.6
-0.8
-1
Notes.
1. For a definition of non-renewable energy consumption see Figure 1.
2. The gray line in the lower panel denotes the 5% significance level.
22
Figure A2 – Rolling regression Granger non-causality tests (Toda and Yamamoto approach)
From renewable energy consumption 1 to real GDP
1
sum of lagged coefficients
p-values of the modified Wald statistic
0.8
0.6
0.4
0.2
0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21
-0.2
From GDP to renewable energy consumption 1
sum of lagged coefficients
0.3
p-values of the modified Wald statistic
0.25
0.2
0.15
0.1
0.05
0
-0.05
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21
-0.1
Notes.
1.
2.
For a definition of renewable energy consumption 1 see Figure 1.
The gray line in the lower panel denotes the 5% significance level.
23
Figure A3 – Rolling regression Granger non-causality tests (Toda and Yamamoto approach)
From renewable energy consumption 2 to real GDP
0.1
0
-0.1 1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
-0.2
sum of lagged coefficients
-0.3
p-values of the modified Wald statistic
-0.4
-0.5
-0.6
-0.7
-0.8
From GDP to renewable energy consumption 2
sum of lagged coefficients
0.9
p-values of the modified Wald statistic
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
Notes.
1. For a definition of renewable energy consumption 2 see Figure 1
2. The gray line in the lower panel denotes the 5% significance level.
24
Table 1 – Energy consumption and economic growth in Italy, average of annual growth
rates in percentages.
1861-1918 1919-1945
1946-1975
1976-2000
Non-renewable energy consumption
2.05
-1.36
6.38
2.44
Renewable energy consumption
3.30
-2.85
12.91
1.16
Renewable and traditional energy
consumption
32.03
4.07
3.98
1.25
Real GDP
0.30
-0.46
0.57
0.39
Note: Author’s calculation on data from Malanima (2006). Data for renewable energy
consumption is from 1887.
25
Table 2 – Zivot and Andrews (1992) unit root tests
Variable
log(real GDP)
∆log(real GDP)
log(NRE)
∆log(NRE)
log(RE1)
∆log(RE1)
log(RE2)
∆log(RE2)
Break
year
Statistic
-3.32
-10.99***
-3.76
-11.31***
-8.56***
-7.47***
-5.60***
-9.50***
1953
1946
1959
1946
1981
1915
1939
1943
Lags
included
in the
model
2
1
1
0
3
3
1
2
Notes
The null of the test is that the series contain a unit root. The 5% critical value of the test is 5.08
and the 1% one is 5.57. Lags were chosen on the basis of the Schwarz criterion.
*** means that the statistic is significant at a 1% level.
26
Table 3 – Granger non-causality tests (Box and Jenkins approach)
From
To
Wald
statistics
0.24
0.32
19.50
0.8
p-values
Sum of lagged
coefficients
-0.02
-0.19
-0.07
-0.01
Causality
Renewable Energy 1 Real GDP
0.88
None
Real GDP
Renewable Energy 1
0.85
None
Renewable Energy 2 Real GDP
0.00
RE2sGDP
Real GDP
Renewable Energy 2
0.96
None
Notes:
1. We adopted a seemingly unrelated regressions model.
2. For a definition of non-renewable energy, renewable energy 1 and renewable energy 2 see
Figure 1
3. In the VAR between renewable energy 1 and real GDP a trend and two mean shifts in 1956
and in 1991 were inserted in the equation for the renewable energy consumption measure on the
basis of specification and stability testing.
27
Table 4 – Unit root and stationarity tests
ADF test
Phillips-Perron test
I
I+T
KPSS test
I
I+T
I
I+T
Variable
log(real
GDP)
0.99(2)
-1.63(2)
0.85
-1.68
0.78*** 0.15**
∆log(real
GDP)
-8.96(1)*** -9.11(1)*** -8.27*** -8.25*** 0.26
0.05
log(NRE)
-0.64(2)
-2.75(1)
-0.53
-2.29
0.79*** 0.08
0.05
∆log(NRE) -8.90(1)*** -8.87(1)*** -9.64*** -9.60*** 0.05
log(RE1)
-13.17(0)*** -8.40(0)*** -21.65*** -28.64*** 0.41*
0.15**
-8.62(0)*** -6.95*** -8.89*** 1.07*** 0.45***
∆log(RE1) -2.92(4)**
0.21
0.06
∆2log(RE1) log(RE2)
-1.80(2)
-4.15(1)*** -1.62
-3.31*
0.39*
0.10
0.03
∆log(RE2) -9.18(2)*** -9.14(2)*** -18.87*** -18.41*** 0.03
Notes:
1. I denotes intercept and I+T denotes intercept and trend
2. ***, ** and * denote significance at 1%, 5% and 10%
3. For the ADF test the number of optimum lags, chosen on the basis of the Schwarz
information criteria, is denoted in parentheses
4. The critical values of the KPSS test are from Kwiatkowski, Phillips, Schmidt and
Shin (1992, Table 1)
5. The critical values for the Phillips and Perron and the ADF tests are based on
MacKinnon (1996) one-sided p-values
6. The Phillips and Perron test is based on the Newey-West bandwidth using a Bartlett
Kernel
7. The KPSS test adopts an Andrews bandwidth using a Bartlett Kernel
8. ∆伊is the first difference operator, while ∆2 is the second difference operator
9. For a definition of NRE, RE1 and RE2 see Figure 1
28
Table 5 – Granger non-causality tests (Toda and Yamamoto approach)
From
To
Modified
Wald
statistics
p-values
Sum of
lagged
coefficients
Causality
Non-renewable
Energy
Real GDP
10.78
0.00
0.05
NREsGDP
Non-renewable
6.48
0.04
-0.76
GDPsNRE
Energy
Renewable Energy 1 Real GDP
0.33
0.84
0.02
None
Renewable Energy
Real GDP
4.72
0.09
0.17
None
1
Renewable Energy 2 Real GDP
19.18
0.00
-0.38
RE2sGDP
Renewable Energy
Real GDP
0.20
0.90
0.04
None
2
Notes:
1. Modified Wald chi-square statistics are displayed.
2. We adopted a seemingly unrelated regressions model after Rambaldi and Doran (1996).
3. For a definition of non-renewable energy, renewable energy 1 and renewable energy 2 see
Figure 1
Real GDP
29
Table 6 – Johansen cointegration tests between the logs of real GDP and non-renewable
energy consumption (138 observations).
Unrestricted Cointegration Rank Test (Trace)
Hypothesized
No. of
Cointegrating
Equations
Eigenvalue
Trace Statistic
0.05 Critical
Value
Prob.**
None
At most 1
0.058991
0.000334
8.436849
0.046155
15.49471
3.841466
0.4199
0.8299
Trace test indicates no cointegration at the 0.05 level
Unrestricted Cointegration Rank Test (Maximum Eigenvalue)
Hypothesized
No. of
Cointegrating
Equations
Eigenvalue
Max-Eigen
Statistic
0.05
Critical Value
Prob.**
None
At most 1
0.058991
0.000334
8.390693
0.046155
14.26460
3.841466
0.3404
0.8299
Max-eigenvalue test indicates no cointegration at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
30
Table 7 – Johansen (1991, 1995) cointegration tests between the logs of real GDP and RE2
consumption (138 observations).
Unrestricted Cointegration Rank Test (Trace)
Hypothesized
No. of
Cointegrating
Equations
Eigenvalue
Trace
Statistic
0.05
Critical Value
Prob.**
None
At most 1
0.072530
0.006156
11.24277
0.852118
15.49471
3.841466
0.1970
0.3560
Trace test indicates no cointegration at the 0.05 level
Unrestricted Cointegration Rank Test (Maximum Eigenvalue)
Hypothesized
No. of
Cointegrating
Equations
None
At most 1
Eigenvalue
Max-Eigen
Statistic
0.05
Critical Value
Prob.**
0.072530
0.006156
10.39065
0.852118
14.26460
3.841466
0.1875
0.3560
Max-eigenvalue test indicates no cointegration at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
31
Table 8 – Summary of the results across different econometric methods.
Variables
under study
Econometric method
Unit root and stationarity
tests of NRE
Cointegration tests
Real GDP
and NRE
VECM
VAR in differences
Real GDP
and RE1
Real GDP
and RE2
Results allowing for
structural breaks
The maximum
integration order is 1
Yes
Bi-directional
causality with
prevailing negative
effects from GDP to
NRE
-
Toda and Yamamoto (1995)
approach
-
Unit root and stationarity
tests of RE1
Cointegration tests
VECM
The maximum
integration order is 0
No
-
VAR in differences
Granger non-causality
Toda and Yamamoto (1995)
approach
Unit root and stationarity
tests of RE2
Cointegration tests
VECM
VAR in differences
Toda and Yamamoto (1995)
approach
Results imposing
parameter stability
The maximum
integration order is 1
No
-
Bi-directional
causality with
prevailing negative
effects from GDP to
NRE
The maximum
integration order is 2
No
Cannot be estimated
given the results of
unit root and
stationarity tests
-
Granger non-causality
The maximum
integration order is 0
No
Negative Granger
causality from RE2 to
real GDP
The maximum
integration order is 1
No
-
-
Negative Granger
causality from RE2 to
real GDP
Notes
Real GDP is always found to be I(1).
32